tanluren

Results 41 comments of tanluren

I try to training custom dataset by using colab.I found some problems with output after training

colab has deactivated GUI ,you can deactivate the relative function of OPENCV or just simply set OPENCV 0 when you make.

No results in evaluation P R and mAP are 0

check the train_batch0.jpg and test_batch0.jpg to see if you prepair the data right

can you provide the pretrain model?

Hi , https://drive.google.com/open?id=1iOsNqQSNanTrbtNYQ5WNSZyxtXHZfjfZ it's trained from scratch and for 150000 batches only,mAP 60 on VOC, more training steps are required. Training with pretrained weights on classification sets would be a...

剪枝时bn层的bias不用mask置零吗？

这一步只是把bn weight置零了，结构还是原来的，bias的激活值也还没转移，只是先测试下初步效果；bias处理在后面的prune_model_keep_size2里，会先将bias的激活偏移值传给后来层，再把它置零。

RuntimeError: Error(s) in loading state_dict for Darknet: size mismatch for module_list.81.Conv2d.weight: copying a param with shape torch.Size([75, 1024, 1, 1]) from checkpoint, the shape in current model is torch.Size([45, 1024, 1, 1]).

the pt does not match the cfg

python -c "from models import *; convert('cfg/yolov3.cfg', 'weights/last.pt')"

是的用自己的cfg和权重

使用同时剪层和通道

已更新这个脚本

请教下关于VOC数据集稀疏化训练的问题

类别多的数据集的确稀疏难度更大，但是也是可以做到的，调小s，用更长的周期去训练

求问yolov4使用哪种剪枝策略比较好

slim_prune.py加layer_prune.py

test

Hi ,which device did you test on ? the speed varies in different device. But it's at least 3x faster than darknet according to my test.